Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Schema as Parameterized Tools for Universal Information Extraction

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Training only 43K token embeddings gives a 1.5B LLM adaptive information extraction across four tasks.

desk verdict A useful parameter-efficient UIE adapter, but the theoretical justification for avoiding schema overlap is wrong and the retrieval metric needs fixing. read the letter →

arxiv 2506.01276 v1 pith:6NL3LVQT submitted 2025-06-02 cs.CL

Sheng Liang , Yongyue Zhang , Yaxiong Wu , Ruiming Tang , Yong Liu This is my paper
classification cs.CL
keywords universalinformationextractionschemaasparameterizedtoolstoolcallingschema-tokenembeddingsparameter-efficientfine-tuningretrievalgenerationon-demand
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes Schema as Parameterized Tools (SPT), a framework that turns information-extraction schemas into callable tokens in a language model's vocabulary. The claim is that training only the schema-token embeddings — roughly 43,000 parameters on a 1.5B model — lets one model retrieve the right schema from a pool, fill its slots, and synthesize a new schema when none fits, unifying closed, open, and on-demand extraction. If correct, adaptive universal information extraction no longer needs full fine-tuning or long in-context demonstrations; extending a schema library becomes a matter of adding token embeddings. Experiments back the central bet: SPT outperforms BM25, dense-retriever, reranker, and finetuned-LoRA baselines at schema selection, and matches LoRA-tuned and leading UIE systems on extraction quality with far fewer trainable parameters.

What carries the argument

The central object is the schema-token embedding matrix $W_S \in \mathbb{R}^{|S|\times d}$, appended to the LLM's vocabulary along with helper tokens <Rej> and <Gen>; the model performs softmax over this extended vocabulary, so an IE schema is selected or generated exactly like a word token. Around this, the framework stacks dual-mode execution — Retrieval Mode (schema retrieval then schema infilling) and Generation Mode (triggered by <Rej>, producing a new schema through <Gen> before infilling) — and a compositional three-phase training strategy that optimizes schema-token alignment first, helper-token behavior second, and joint tuning last. This machinery is what lets a frozen or lightly adapted LLM retrieve, create, and fill schemas in one unified token space.

What would settle it

Build a schema pool with overlapping or semantically conflicting schemas (for instance, two schemas that share slot names but use different types, or the same event type under different names) and train SPT on it; if schema retrieval accuracy or extraction F1 drops sharply relative to the non-overlapping case, or if adding schemas to the pool degrades retrieval of earlier schemas, the Schema Independence assumption fails. Concretely, also measure retrieval Recall@k as the pool grows from 26 toward hundreds of schemas; a plateau or decline would falsify the scaling claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the LLM's tool-calling mechanism is a natural home for universal information extraction: predefined schemas can be reimagined as parameterized tools, so extraction becomes tool selection plus parameter filling. SPT extends the model's embedding matrix with one trainable token per schema plus two helper tokens, <Rej> and <Gen>. During inference the model either retrieves a matching schema token and infills its arguments, or emits <Rej> to switch into generation mode and synthesizes a new schema via <Gen>. A three-phase compositional training schedule first aligns schema tokens with closed-schema extractions, then teaches the helper tokens, then jointly fine-tunes all of them. Reported results show strong schema retrieval across the unified 26-schema pool and extraction F1 scores comparable to LoRA and leading UIE baselines, with only about 43K trainable parameters against LoRA's roughly 1.2M.

Load-bearing premise

The paper assumes 'Schema Independence' — that because its 26-schema benchmark is deliberately non-overlapping, each schema token's training signal stays decoupled and retrieval performance remains robust to accuracy fluctuations, so the framework will scale to numerous schemas in realistic settings.

Editorial extensions

If this is right

  • Schema libraries become pluggable: adding a new IE task reduces to training or appending a new schema-token embedding rather than re-tuning the model.
  • IE systems can offer governance-friendly structured output while still covering unseen cases through on-the-fly schema generation, since <Rej> and <Gen> give an explicit fallback path.
  • Because retrieval is learned as generation rather than semantic similarity, SPT can select schemas whose descriptions would not surface via BM25 or dense retrievers, as shown on CrudeOilNews (0.76 vs 0.52 best baseline).
  • The parameter count of around 43K makes adaptive UIE feasible in low-resource settings where LoRA's 1.2M trainable parameters or full fine-tuning is impractical.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's 'Schema Independence' assumption is untested for overlapping schemas; a natural extension is to construct a pool with intentionally conflicting or redundant schemas and check whether retrieval and infilling degrade, since real-world schema libraries rarely partition cleanly.
  • The finding that a smaller model with untied input/output embeddings can rival a larger tied-embedding model suggests that embedding tiedness is a design lever for token-only tuning, and could be tested systematically across architectures.
  • SPT's token-as-tool view may transfer to other structured-generation domains, such as API selection with parameter filling or database query construction, where the same retrieve-or-generate-and-fill loop applies.
  • Retrieval-mode robustness is measured on a small 26-schema pool; scaling the pool toward hundreds of schemas is the obvious next experiment to see whether token embeddings saturate or interfere.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Schema as Parameterized Tools (SPT), a universal information extraction framework in which predefined extraction schemas are treated as trainable token embeddings appended to the LLM vocabulary, together with <Rej> and <Gen> helper tokens for switching between retrieval of existing schemas and on-the-fly schema generation. The method is evaluated on four IE tasks (NER, RE, EE, ODIE) using Qwen2.5-1.5B and other small backbones, with only about 43K trainable parameters. The reported results show strong schema retrieval relative to sparse/dense retrieval baselines and extraction F1 comparable to LoRA baselines on several datasets.

Significance. If the empirical claims hold, the paper makes a useful contribution: it shows that a single set of schema-token embeddings can serve as a parameter-efficient adapter for closed, open, and on-demand IE, and the cross-backbone ablation is a nice check of generality. The paper is generally clearly written and the limitation section is candid about the lack of overlapping-schema benchmarks and large-scale validation. The main novelty—treating schemas as parameterized tokens, analogous to toolken methods—is a reasonable extension of prior work, and the reported parameter efficiency is striking. However, the central scalability claim rests on an incorrect theoretical assertion about gradient decoupling, and the retrieval comparison in Table 1 is not apples-to-apples, so the strength of the conclusions is currently overstated.

major comments (3)
  1. [§4.1] The 'Schema Independence' claim is mathematically incorrect for the softmax cross-entropy objective defined in Section 3.2. For a sample with gold schema token a, the gradient with respect to another schema embedding b (b ≠ a) is dL/dw_b = P(b|x) h, where h is the hidden state before the output projection. Because the softmax normalizes over the entire extended vocabulary, P(b|x) is strictly positive almost everywhere; non-overlap of training samples across tasks does not make it zero. Thus gradients for schema-token embeddings are coupled through every sample. The paper's assertion that non-overlapping task allocation 'ensures stable training even with schema scaling' is therefore unsupported. Since the unified benchmark deliberately contains only 26 non-overlapping schemas, the experiments validate the method only in the regime where cross-schema interference cannot occur. Please either remove the decoupling claim and reframe it as an untested assumption, or add experiments with overlapping/conflicting schemas and report retrieval/extraction behavior under interference.
  2. [§5.1, Table 1] The schema-retrieval comparison is not controlled. BM25, BGE-M3, and BGE-Reranker are scored with Recall@k=5, while SPT and the fine-tuned generator produce a variable number of schema tokens, and the paper does not report the average number of generated schemas per query. If SPT emits more than five schema candidates on average, higher recall is expected by construction. Please report the average k for the generative methods, or evaluate all methods with the same controlled k, and provide per-query variance. Without this, the 'robust schema retrieval and selection performance' claim is overstated.
  3. [§5.1–§5.3] No error bars, confidence intervals, or multiple-seed results are reported anywhere. Several headline differences are small (e.g., Table 3 header F1 0.69 vs. 0.71; Table 1 CrudeOilNews retrieval 0.76 vs. 0.52), and with only one run it is impossible to assess whether the improvements are significant. Please add at least 3–5 seeds with mean and std for the main comparisons, or otherwise justify why single runs are sufficient for the claims of superiority and comparability.
minor comments (5)
  1. [§5.3] In the paragraph after Table 3, the text reads 'e donot report a zeroshot baseline because the difficulty of the task is too high for a 1.5B pretraind model' and later 'ex- treamly low parameter size'; these should be corrected to 'We do not report', 'pretrained', and 'extremely'.
  2. [§4.1] The sentence 'Thus we assume our extraction performance will maintain robustness despite retrieval accuracy may variations' is ungrammatical and should be rewritten, e.g., 'We assume that extraction performance will remain robust despite possible variations in retrieval accuracy.'
  3. [§5.2, Table 2] In the 'Finetuned w/ Gold Schemas' row, the dash for the two 'Reject' columns should be explained in the table caption or text: rejection is not applicable when gold schemas are always provided, rather than being a missing value.
  4. [§4.2] There is a typo in the baseline description: 'Fituning' should be 'Finetuned'.
  5. [§4.1] The term 'ToolenGPT' in the sentence 'unlike ToolenGPT with 200+ tools' should be 'ToolkenGPT' for consistency with the related-work citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SPT is an empirical supervised-learning paper whose reported retrieval/infilling/generation numbers are held-out test measurements, not fitted values relabeled as predictions.

full rationale

No step in the paper is a derivation that reduces by construction to its own inputs. SPT is an empirical fine-tuning method: the 28 trainable schema/helper embeddings are optimized on training data, and the reported retrieval F1, extraction F1, and ROUGE-L scores are computed on held-out test splits, so the results are genuine out-of-sample measurements rather than fitted parameters relabeled as predictions. The paper does not fit a model to a subset and then 'predict' the same subset in disguised form. The §4.1 'Schema Independence' passage asserts that gradients for each schema token are decoupled due to non-overlapping task allocation; this is an unproven and mathematically questionable assumption under softmax cross-entropy, but it is not equivalent to, nor derived from, the paper's measured outcomes, so it is a correctness or robustness concern rather than circularity. Similarly, the limitations stated in §7 explicitly concede the restricted evaluation scale and the absence of overlapping-schema protocols; these caveats weaken generalizability claims but do not create a self-referential reduction. The self-citations present in the paper (e.g., ToolACE in Related Work) are contextual and not load-bearing; no central claim is justified solely by the authors' prior work, and no uniqueness theorem is imported to force the method choice. The central novelty, schema tokens for adaptive IE, is evaluated against external baselines on standard datasets, so the claim of 'robust schema retrieval and selection performance' has independent empirical content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim rests on learned schema and helper token embeddings, plus hand-chosen data and evaluation settings. The Schema Independence assumption is explicitly stated but not proven. No physical or external entities are introduced.

free parameters (5)
  • Schema token embedding matrix WS (26 x 1536) = learned on CrudeOilNews, SciERC, AnatEM training data
    Each of the 26 schema tokens is a trainable embedding optimized in Phases 1 and 3; retrieval and infilling depend directly on these vectors. This is a fitted parameter, not a derived constant.
  • Helper token embeddings w<Rej>, w<Gen> = learned
    The <Rej> and <Gen> tokens control the dual-mode decision; their embeddings are trained in Phases 2 and 3 and are central to the claimed adaptive behavior.
  • Downsampling ratio for schema-free samples = 30%
    Schema-free training samples were downsampled to 30% to balance retrieval and generation learning (Section 4.1). This is a hand-chosen data balance that can affect rejection and generation behavior and final F1.
  • Top-k for retrieval evaluation = 5
    Recall@k=5 is the chosen operating point for retrieval baselines; generative methods are allowed variable k, which affects comparability.
  • Training hyperparameters (LR 5e-4, 5e-5; 3+2 epochs) = stated in Section 4.3
    Learning rates and epoch counts are hand-set and affect convergence and final performance.
assumptions (4)
  • standard math Softmax next-token distribution over the extended vocabulary V∪S∪{<Rej>,<Gen>} is a valid model of schema selection and infilling.
    Section 3.2 defines P(x_i|x<i)=softmax(W·h_{i-1}); this is the standard LM head formalism, used without proof.
  • domain assumption Adding trainable token embeddings to the vocabulary does not degrade the base LLM's existing reasoning or extraction abilities.
    Section 3.2 cites Wang et al. 2024 for this; the paper relies on it to claim the 1.5B backbone remains capable after embedding-only tuning.
  • domain assumption Schema Independence: non-overlapping schema sets decouple gradients per schema token, enabling stable training and performance robustness with schema scaling.
    Section 4.1 states this assumption explicitly and uses it to justify restricting evaluation to 26 non-overlapping schemas; if false, the massive-schema claim is unsupported.
  • domain assumption ODIE soft-matching F1 and ROUGE-L scores measure schema generation quality adequately.
    Section 4.2 and Table 3 adopt the ODIE evaluation protocol without re-validating its reliability.
invented entities (2)
  • <Rej> helper token
    purpose: Signals that no predefined schema matches the query and triggers schema generation.
    A new trainable token embedding introduced by SPT; its behavior is learned from the authors' data and has no falsifiable prediction outside this paper's evaluation.
  • <Gen> helper token
    purpose: Acts as a pseudo schema token representing newly generated schemas during on-the-fly generation.
    A new trainable token embedding introduced by SPT; its effectiveness is only measured on the ODIE benchmark used to train it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Schema as Parameterized Tools for Universal Information Extraction." pith.science (2026). https://pith.science/paper/6NL3LVQT

@misc{pith2026250601276,
  author       = {Pith},
  title        = {Pith review of: Schema as Parameterized Tools for Universal Information Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NL3LVQT}},
  note         = {Machine review of arXiv:2506.01276}
}
read the original abstract

Universal information extraction (UIE) primarily employs an extractive generation approach with large language models (LLMs), typically outputting structured information based on predefined schemas such as JSON or tables. UIE suffers from a lack of adaptability when selecting between predefined schemas and on-the-fly schema generation within the in-context learning paradigm, especially when there are numerous schemas to choose from. In this paper, we propose a unified adaptive text-to-structure generation framework, called Schema as Parameterized Tools (SPT), which reimagines the tool-calling capability of LLMs by treating predefined schemas as parameterized tools for tool selection and parameter filling. Specifically, our SPT method can be applied to unify closed, open, and on-demand IE tasks by adopting Schema Retrieval by fetching the relevant schemas from a predefined pool, Schema Filling by extracting information and filling slots as with tool parameters, or Schema Generation by synthesizing new schemas with uncovered cases. Experiments show that the SPT method can handle four distinct IE tasks adaptively, delivering robust schema retrieval and selection performance. SPT also achieves comparable extraction performance to LoRA baselines and current leading UIE systems with significantly fewer trainable parameters.

Figures

Figures reproduced from arXiv: 2506.01276 by the authors.

Figure 1
Figure 1. An overview of UIE. et al., 2007; Fader et al., 2011; Stanovsky et al., 2018), and (3) On-demand IE where extraction targets are dynamically specified through natural language instructions (Jiao et al., 2023). UIE has demonstrated superior schema adaptability com￾pared to traditional IE systems (Li et al., 2023) that are tailored for specific tasks such as named entity recognition (NER), relation extraction (RE), an… view at source ↗
Figure 2
Figure 2. Overview of our proposed Schema as Parameterized Tools (SPT) framework. Schema-token embeddings [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. , 6 and 7 show examples of LLM retrieval, extraction, and generation, respectively. B Time Complexity [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: SPT retrieval mode [PITH_FULL_IMAGE:figures/full_fig_p012_3.png]
Figure 6
Figure 6. Figure 6: LLM extraction. LLM generation. [Prompt] You are a helpful assistant. Generate schemas to extract information. [Input] John, a professor at MIT, founded TechCorp in 2010. [Output] Based on the query, I should extract: <Organization> Extraction results: {‘Organization’:…
Figure 4
Figure 4. Figure 4: SPT generation mode. A.2 LLM [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 21 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    David Ahn. 2006. https://aclanthology.org/W06-0901/ The stages of event extraction . In Proceedings of the Workshop on Annotating and Reasoning about Time and Events, pages 1--8, Sydney, Australia. Association for Computational Linguistics

    work page 2006

  4. [4]

    Cafarella, Stephen Soderland, Matthew Broadhead, and Oren Etzioni

    Michele Banko, Michael J. Cafarella, Stephen Soderland, Matthew Broadhead, and Oren Etzioni. 2007. https://api.semanticscholar.org/CorpusID:207169186 Open information extraction from the web . In CACM

    work page 2007

  5. [5]

    Tianle Cai, Xuezhi Wang, Tengyu Ma, Xinyun Chen, and Denny Zhou. 2024. https://openreview.net/forum?id=qV83K9d5WB Large language models as tool makers . In The Twelfth International Conference on Learning Representations

    2024

  6. [6]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. https://arxiv.org/abs/2402.03216 Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation . Preprint, arXiv:2402.03216

    arXiv 2024

  7. [7]

    Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. 2023. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220--235

    2023

  8. [8]

    Sefika Efeoglu and Adrian Paschke. 2024. Retrieval-augmented generation-based relation extraction. arXiv preprint arXiv:2404.13397

    arXiv 2024

Show all 52 references
  1. [9]

    Anthony Fader, Stephen Soderland, and Oren Etzioni. 2011. https://aclanthology.org/D11-1142/ Identifying relations for open information extraction . In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1535--1545, Edinburgh, Scotland...

    2011

  2. [10]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

    2023 arXiv

  3. [11]

    Honghao Gui, Lin Yuan, Hongbin Ye, Ningyu Zhang, Mengshu Sun, Lei Liang, and Huajun Chen. 2024. https://doi.org/10.18653/v1/2024.acl-short.13 IEP ile: Unearthing large scale schema-conditioned information extraction corpus . In Proceedings of the 62nd Annual Meeting of the Ass...

    2024 doi

  4. [12]

    Yucan Guo, Zixuan Li, Xiaolong Jin, Yantao Liu, Yutao Zeng, Wenxuan Liu, Xiang Li, Pan Yang, Long Bai, Jiafeng Guo, et al. 2023. Retrieval-augmented code generation for universal information extraction. arXiv preprint arXiv:2311.02962

    2023 arXiv

  5. [13]

    Xu Han, Tianyu Gao, Yankai Lin, Hao Peng, Yaoliang Yang, Chaojun Xiao, Zhiyuan Liu, Peng Li, Jie Zhou, and Maosong Sun. 2020. https://doi.org/10.18653/v1/2020.aacl-main.75 More data, more relations, more context and more openness: A review and outlook for relation extraction ....

    2020 doi

  6. [14]

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. 2024. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608

    2024 arXiv

  7. [15]

    Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. 2023. https://openreview.net/forum?id=BHXsb69bSx Toolken GPT : Augmenting frozen language models with massive tools via tool embeddings . In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  8. [16]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  9. [17]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  10. [18]

    Yizhu Jiao, Ming Zhong, Sha Li, Ruining Zhao, Siru Ouyang, Heng Ji, and Jiawei Han. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.620 Instruct and extract: Instruction tuning for on-demand information extraction . In Proceedings of the 2023 Conference on Empirical Methods ...

    2023 doi

  11. [19]

    Meisin Lee, Lay-Ki Soon, Eu Gene Siew, and Ly Fie Sugianto. 2022. https://aclanthology.org/2022.lrec-1.49/ C rude O il N ews: An annotated crude oil news corpus for event extraction . In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 465--479...

    2022

  12. [20]

    Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691

    2021 arXiv

  13. [21]

    Bo Li, Gexiang Fang, Yang Yang, Quansen Wang, Wei Ye, Wen Zhao, and Shikun Zhang. 2023. Evaluating chatgpt's information extraction capabilities: An assessment of performance, explainability, calibration, and faithfulness. arXiv preprint arXiv:2304.11633

    2023 arXiv

  14. [22]

    Qi Li, Heng Ji, and Liang Huang. 2013. https://aclanthology.org/P13-1008/ Joint event extraction via structured prediction with global features . In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 73--82, S...

    2013

  15. [23]

    Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190

    2021 arXiv

  16. [24]

    Zixuan Li, Yutao Zeng, Yuxin Zuo, Weicheng Ren, Wenxuan Liu, Miao Su, Yucan Guo, Yantao Liu, Lixiang Lixiang, Zhilei Hu, Long Bai, Wei Li, Yidan Liu, Pan Yang, Xiaolong Jin, Jiafeng Guo, and Xueqi Cheng. 2024. https://doi.org/10.18653/v1/2024.acl-long.475 K now C oder: Coding ...

    2024 doi

  17. [25]

    Weiwen Liu, Xingshan Zeng, Xu Huang, xinlong hao, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Zhengying Liu, Yuanqing Yu, Zezhong WANG, Yuxian Wang, Wu Ning, Yutai Hou, Bin Wang, Chuhan Wu, Wang Xinzhi, Yong Liu, Yasheng Wang, Duyu Tang, Dandan Tu, Lifeng Shang, Xin Jiang, Rui...

    2025

  18. [26]

    Yaojie Lu, Qing Liu, Dai Dai, Xinyan Xiao, Hongyu Lin, Xianpei Han, Le Sun, and Hua Wu. 2022. https://doi.org/10.18653/v1/2022.acl-long.395 Unified structure generation for universal information extraction . In Proceedings of the 60th Annual Meeting of the Association for Comp...

    2022 doi

  19. [27]

    Yi Luan, Luheng He, Mari Ostendorf, and Hannaneh Hajishirzi. 2018. https://doi.org/10.18653/v1/D18-1360 Multi-task identification of entities, relations, and coreference for scientific knowledge graph construction . In Proceedings of the 2018 Conference on Empirical Methods in...

    2018 doi

  20. [28]

    David McClosky, Mihai Surdeanu, and Christopher Manning. 2011. https://aclanthology.org/W11-1806/ Event extraction as dependency parsing for B io NLP 2011 . In Proceedings of B io NLP Shared Task 2011 Workshop , pages 41--45, Portland, Oregon, USA. Association for Computationa...

    2011

  21. [29]

    Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. 2009. https://aclanthology.org/P09-1113/ Distant supervision for relation extraction without labeled data . In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Co...

    2009

  22. [30]

    Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grishman. 2016. https://doi.org/10.18653/v1/N16-1034 Joint event extraction via recurrent neural networks . In Proceedings of the 2016 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human L...

    2016 doi

  23. [31]

    Chaoxu Pang, Yixuan Cao, Qiang Ding, and Ping Luo. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.950 Guideline learning for in-context information extraction . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15372--15389, Si...

    2023 doi

  24. [32]

    Jonas Pfeiffer, Andreas R \"u ckl \'e , Clifton Poth, Aishwarya Kamath, Ivan Vuli \'c , Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020. Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779

    2020 arXiv

  25. [33]

    Sampo Pyysalo and Sophia Ananiadou. 2014. https://doi.org/10.1093/BIOINFORMATICS/BTT580 Anatomical entity mention recognition at literature scale . Bioinform., 30(6):868--875

    2014 doi

  26. [34]

    Yunjia Qi, Hao Peng, Xiaozhi Wang, Bin Xu, Lei Hou, and Juanzi Li. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.419 ADELIE : Aligning large language models on information extraction . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processin...

    2024 doi

  27. [35]

    Cheng Qian, Chenyan Xiong, Zhenghao Liu, and Zhiyuan Liu. 2024. https://doi.org/10.18653/v1/2024.naacl-long.48 Toolink: Linking toolkit creation and using through chain-of-solving on open-source model . In Proceedings of the 2024 Conference of the North American Chapter of the...

    2024 doi

  28. [36]

    Robertson and Hugo Zaragoza

    Stephen E. Robertson and Hugo Zaragoza. 2009. https://doi.org/10.1561/1500000019 The probabilistic relevance framework: BM25 and beyond . Found. Trends Inf. Retr., 3(4):333--389

    2009 doi

  29. [37]

    Oscar Sainz, Iker Garc \' a-Ferrero, Rodrigo Agerri, Oier Lopez de Lacalle, German Rigau, and Eneko Agirre. 2024. https://openreview.net/forum?id=Y3wpuxd7u9 Go LLIE : Annotation guidelines improve zero-shot information-extraction . In The Twelfth International Conference on Le...

    2024

  30. [38]

    Tjong Kim Sang and Fien De Meulder

    E. Tjong Kim Sang and Fien De Meulder. 2003. https://api.semanticscholar.org/CorpusID:2470716 Introduction to the conll-2003 shared task: Language-independent named entity recognition . In Conference on Computational Natural Language Learning

    2003

  31. [39]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dess \` , Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36:68539--68551

    2023

  32. [40]

    Fatemeh Shiri, Van Nguyen, Farhad Moghimifar, John Yoo, Gholamreza Haffari, and Yuan-Fang Li. 2024. Decompose, enrich, and extract! schema-aware event extraction using llms. arXiv preprint arXiv:2406.01045

    2024 arXiv

  33. [41]

    Gabriel Stanovsky, Julian Michael, Luke Zettlemoyer, and Ido Dagan. 2018. https://doi.org/10.18653/v1/N18-1081 Supervised open information extraction . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human ...

    2018 doi

  34. [42]

    Renxi Wang, Xudong Han, Lei Ji, Shu Wang, Timothy Baldwin, and Haonan Li. 2024. Toolgen: Unified tool retrieval and calling via generation. arXiv preprint arXiv:2410.03439

    2024 arXiv

  35. [43]

    Xinglin Xiao, Yijie Wang, Nan Xu, Yuqi Wang, Hanxuan Yang, Minzheng Wang, Yin Luo, Lei Wang, Wenji Mao, and Daniel Zeng. 2023. Yayi-uie: A chat-enhanced instruction tuning framework for universal information extraction. arXiv preprint arXiv:2312.15548

    2023 arXiv

  36. [44]

    Derong Xu, Wei Chen, Wenjun Peng, Chao Zhang, Tong Xu, Xiangyu Zhao, Xian Wu, Yefeng Zheng, Yang Wang, and Enhong Chen. 2024 a . Large language models for generative information extraction: A survey. Frontiers of Computer Science, 18(6):186357

    2024

  37. [45]

    Lingling Xu, Haoran Xie, Si-Zhao Joe Qin, Xiaohui Tao, and Fu Lee Wang. 2023. Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148

    2023 arXiv

  38. [46]

    Qiancheng Xu, Yongqi Li, Heming Xia, and Wenjie Li. 2024 b . https://doi.org/10.18653/v1/2024.findings-emnlp.561 Enhancing tool retrieval with iterative feedback from large language models . In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9609--...

    2024 doi

  39. [47]

    Vikas Yadav and Steven Bethard. 2018. https://aclanthology.org/C18-1182/ A survey on recent advances in named entity recognition from deep learning models . In Proceedings of the 27th International Conference on Computational Linguistics, pages 2145--2158, Santa Fe, New Mexico...

    2018

  40. [48]

    Konstantin Yakovlev, Sergey Nikolenko, and Andrey Bout. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.345 Toolken+: Improving LLM tool usage with reranking and a reject option . In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5967--5974,...

    2024 doi

  41. [49]

    Lifan Yuan, Yangyi Chen, Xingyao Wang, Yi Fung, Hao Peng, and Heng Ji. 2024. https://openreview.net/forum?id=G0vdDSt9XM CRAFT : Customizing LLM s by creating and retrieving from specialized toolsets . In The Twelfth International Conference on Learning Representations

    2024

  42. [50]

    Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. 2021. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199

    2021

  43. [51]

    Yuanhang Zheng, Peng Li, Wei Liu, Yang Liu, Jian Luan, and Bin Wang. 2024. https://aclanthology.org/2024.lrec-main.1413/ T ool R erank: Adaptive and hierarchy-aware reranking for tool retrieval . In Proceedings of the 2024 Joint International Conference on Computational Lingui...

    2024

  44. [52]

    Zexuan Zhong and Danqi Chen. 2021. https://doi.org/10.18653/v1/2021.naacl-main.5 A frustratingly easy approach for entity and relation extraction . In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2021 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.